Things I learned at SSC2022

5: Clustering Pareto Fronts

Y. Goto, H. Morita, Y. Shirai, and H. Ichikawa;
Simulation-based classification in multi-objective optimization problems with social simulation


  1. You have an ABM
  2. You can generate a pareto front to trade-off multiple objectives
  3. But there are multiple scenarios (prefectures, in this case)


It would be nice if we could draw some insight on the general tradeoffs faced across scenarios

5: Clustering Pareto Fronts


Generate the pareto fronts, then let’s cluster them somehow.


How do you cluster pareto fronts?

  1. Similarity of decision variables
  2. Similarity of model objectives

5: Clustering Pareto Fronts

 

4: Neural Network/RL and its limits

Thomas Chesney, “A philosophy of intelligent agent-based models

 


We can use reinforcement learning with neural networks to teach cars how to behave

4: Neural Network/RL and its limits

For the right amount of sophistication, they eventually learn better behaviour.


The problem is, what does a traffic model with super-smart agents actually means?


ABMs are often not operational puzzles to maximize; agent behaviour needs to be proportional to the problem at hand.


Even though we can, often we shouldn’t.

3: Inverse Generative Social Science

Lux Miranda; “Evolutionary model discovery of human behavioral factors driving decision-making in irrigation experiments


  1. Public goods game, why do people contribute?
  2. Lab experiment with undergrads
  3. What are their utility functions?


Instead of writing a utility function and calibrating, let’s use genetic programming to evolve the utility function itself.

Inverse Generative Social Science


 

Inverse Generative Social Science


It works as a gentle introduction to inverse generativve social science.


When the problem and functions are this simple however, IGS is really a “calibration with genetic programming”

2: The dangers of random movement

Edmund Chattoe-Brown , “All the right moves? Systematically exploring the effects of random travel in agent-based models



Not the only way to be random; this method produces “line of sight” explorations.

2: The dangers of random movement


Default Movement

New Movement

“Movement” in the model is a complicated proxy for habitat and exploration. Without estimation/calibration, every component of an abstract model matters.

1. The march of data-assimilation

 

1. The march of data-assimilation


Alison Heppenstall; “Simulating social systems with individual-based models: are they worth it?


Deborah Olukan, Jonathan Ward, Nicolas Malleson & Jiaqi Ge; “Heterogeneity in agent-based models


Joel Dyer; “Black-box Bayesian inference for agent-based models